REDISTRIBUTIVE TAXATION AS SOCIAL INSURANCE

Transcription

REDISTRIBUTIVE TAXATION AS SOCIAL INSURANCE
Journal
of Public
Economics
14 (1980) 49-68
REDISTRIBUTIVE
0 North-Holland
TAXATION
Publishing
AS SOCIAL
Company
INSURANCE
Hal R. VARIAN*
University
Received
of Michigan,
October
Ann Arbor, MI 48109, USA
1979, revised version
received
March
1980
The modern literature on nonlinear optimal taxation treats differences in income as being due to
unobserved
differences in ability. A striking result of this assumption
is that high income agents
should face a zero marginal tax rate. In this paper I assume that differences in observed income
are due to exogenous differences in luck. Hence the optimal redistributive
tax involves trading
off the benefits due to ‘social insurance’ with the costs due to reduced incentives. I derive the
some algebraic
and numeric
optimal
forms for linear and nonlinear
taxes, and compute
examples. Typically high income individuals will face quite high marginal tax rates.
1. Introduction
Recent work on the sources of income inequality
has found that only a
small fraction
of the variation
in income can be explained
by observed
socioeconomic
characteristics.
For example, after an extensive examination
of
the available evidence, Jencks and his co-workers
conclude:
‘Neither family
background,
cognitive skill, educational
attainment,
nor occupational
status
explains much of the variation
in men’s income. Indeed when we compare
men who are identical in all these respects, we find only 12 to 15 percent less
inequality than among random individuals’
[Jencks (1972, p, 266)].
Jencks goes on to list three possible causes for such unexplained
variations
in income.
(1) Differences in tastes: ‘. . . some men value money more than others, and
these men make unusual sacrifices to get it’.
(2) Differences in unobserved
endowments
of native ability: ‘. . . the ability
to hit a ball thrown at high speed, the ability to type a letter quickly and
accurately, the ability to persuade a customer that he wants a larger car than
he thought he wanted.. . and so forth’.
(3) Differences in luck: ‘. . chance acquaintances
who steer you to one line
of work rather than another, the range of jobs that happens to be available
in a particular
community
when you are job hunting,
the amount
of
*I have benefited from comments
by Theodore
Bergstrom,
Lamberto
Cesari.
Bengt
Holmstriim,
Glenn Loury, James Mirrlees, Carl Simon and the anonymous
referees of this
journal. I am of course responsible for any remaining errors. This material is based upon work
supported by the National Science Foundation
under Grant SOC78-05757
and the Guggenheim
Memorial Foundation.
H.R. Varian, Redistrihutiue
50
taxation
overtime work in your particular
plant, whether bad weather destroys your
accidents’.
strawberry crop.. . and a hundred other unpredictable
Jencks and his co-workers argue that this last factor is probably the most
important
single influence on the distribution
of income. It is difficult to
support this view by appealing only to cross-sectional
data since one cannot
adequately
control
for variations
in unobserved
tastes and endowments.
However, other econometric
studies of income determination
using panel
data have tended to confirm Jencks’s view. Coe (1977) for example, describes
some interesting
figures which were extracted
from the University
of
Michigan income dynamics study of 5000 American families:
Of the individuals
who were poor in the one year period 1975, only 12
percent were in poverty in every one of the nine years between 1967 and
1975. On the other hand, while only 8.9 percent of the population
was
poor in 1975, fully one-quarter
of the sample individuals
were in poverty
in at least one of the nine years between 1967 and 1975.
Similar conclusions
were reached by Lillard and Willis (1978) who used the
University of Michigan income dynamics data for the years 1967-73:
While the poor are different in the sense just described,
it would be
misleading to conclude that poverty is a permanent
status. We find that of
those individuals
in poverty in a given year, about 55 percent of the
whites and 35 percent of the blacks will be out of poverty in the following
year. Another indication
of mobility is that only 15 percent of whites and
35 percent of blacks who fall into poverty sometime during the three year
period from 1967-70 are expected to be in poverty in all three years (p.
1007).
These data suggest that the movements
of individual
incomes over time
contain a large random component,
i.e. a component
that is not explained by
differences in tastes and endowments.
Such a view ‘tends to modify attitudes concerning
policy towards income
distribution,
and in particular, policies concerned with redistributive
taxation.
Most discussions
of redistributive
taxation
have taken place in a context
where uncertainty
was either ignored, or where markets for transferring
risk
were sufficiently well developed so that no risk was borne unintentionally.
The above empirical evidence shows that randomness
in income is a major
problem
which cannot
simply be ignored,
and well-known
arguments
concerning
moral hazard, adverse selection, transactions
costs, and returns to
scale show why complete markets for shifting risks may be unavailable.
Let us consider then how adding a random component
to income may
affect the analysis of redistributive
programs.
Suppose that we consider a
simple world
Here r
of identical
is the expected
risk-averse
income
individuals
of each individual
each with income
x =Y+si.
and ci is a random
error
H.R. Varian, Redistributive
taxution
51
term assumed to be uncorrelated
between individuals,
and to have expected
value of zero. Since ai is uncorrelated
between individuals,
there is a clear
economic
case for the establishment
of an insurance
market which could
eliminate individual
risk. However, let us suppose that effects such as those
described above are present so that an insurance
market cannot be viable.
Then there is still a simple government
policy to improve welfare: impose a
100 percent tax on observed income and redistribute
a uniform grant r Such
a program of redistributive
taxation would effectively replace the nonexistent
insurance
market and provide an unambiguous
increase in welfare. The
motive for redistribution
here is not a desire for equity per se, but rather a
desire for social insurance.
The above points are couched in somewhat
abstract language, but it is
clear that the basic idea is commonplace.
Indeed, the fact that redistributive
taxation helps to insure against individual
risk is a common justification
for
redistributive
programs.
Consider,
for example, the following excerpt from
the Beveridge Report, a study which provided
the guidelines
for British
welfare policy in the post-war period [Beveridge (1942)]:
Abolition
of want cannot
be brought
about
merely
by increasing
production,
without seeing to correct distribution
of the product., . . Better
is required
among
wage earners
distribution
of purchasing
power
themselves, as between times of earning and not earning, and between times
of heavy fumily responsibilities
and light or no family responsibilities.
Both
social insurance
and children’s
allowances
are primary
methods
of
redistributing
wealth. (Emphasis added.)
Similar
sentimates
are expressed
by Stanford
G. Ross, the current
Commissioner
of U.S. Social Security: ‘Everybody is vulnerable
to a sudden
reduction
in income, regardless of his or her station in life. Society needs
some mechanism to guard against the risk when it is most likely to occur.. .”
[Ross (1979)].
Jencks and his co-workers
have suggested that the most effective way to
provide a more equal distribution
of income would be to engage in a
widespread system of redistributive
taxation. The above discussion indicates
that even apart from the possibly
desirable
effect of reducing
income
inequality,
there is a definitely desirable effect of improving
the allocation
of
risk bearing. Indeed, I suspect that widespread
political support of many
redistributive
programs
rests more with the social insurance
aspect of the
program than with altruistic consideration
involving social welfare.
Of course, there is a cost to any program of redistributive
taxation, namely
the incentive loss that is involved in any tax or transactions.
The choice of
an optimal redistributive
tax therefore involves trading off three kinds of
effects: (1) the equity, effect of changing
the distribution
of income; (2) the
52
H.R. Varian, Redistributiue
taxation
and (3) the insurance effect from
efficiency effect from reducing incentives;
reducing the variance of individual
income streams.
The first two aspects have been examined extensively in recent years. Some
relevant articles are those by Atkinson
and Stiglitz (1976), Diamond
(1975),
Diamond
and Mirrlees (1971), and Mirrlees (1971). Sandmo (1976) provides
a brief survey with bibliographic
references.
The insurance effect of income redistribution
has received considerably
less
attention in the theoretical literature.
A notable exception has been the work of Mirrlees and his collaborators.
Mirrlees (1974) posed the problem of the insurance-incentive
tradeoff in a
fairly general model and derived
a formula
characterizing
the optimal
(second-best)
policy. Subsequently,
Diamond
and Mirrlees (1978) analyzed a
model of a Social Security
type program,
emphasizing
the insuranceincentive aspects involved in the retirement
decision. Diamond,
Helms and
Mirrlees (1978) present some numerical examples of social insurance benefits
in some simple cases involving linear taxation.
In this paper I formulate and analyze a simple model of social insurance.
Section
2 describes
the basic model. In section
3 I derive a formula
characterizing
the optimal linear tax for this model. Section 4 is devoted to
an algebraic and numerical example of the linear tax. Section 5 describes the
optimal nonlinear
tax, and section 7 provides a brief summary. An appendix
describes an existence theorem for a class of optimization
problems relevant
to social insurance problems.
2. A problem in social insurance
Let us consider a simple model involving social insurance. We have a large
number of identical consumers who wish to transfer some amount of wealth
(w) between two time periods. We let x be the amount saved in the first time
period so that consumption
at that time is w-x. Consumption
in the second
period will be x +E, where E is a random
variable with mean zero. The
random
variable
E is supposed
to represent
the cumulative
effect of
exogenous
‘luck’ on the individual
consumer.
We suppose
luck to be
independently
and identically distributed
among the individual
agents. Since
luck is assumed to be independently
distributed
from consumer to consumer,
there is a clear market incentive to provide insurance.
In this framework, the individual
optimization
problem is:
maxu(w-x)+Eu(x+s),
X
(1)
where u(y) is a von Neuman-Morgenstern
utility function
and E is the
expectation
operator. We assume that u(y) exhibits the standard properties
of monotonicity
in income and risk aversion. At times we will assume that
the utility function
exhibits declining
absolute
risk aversion.
These three
behavioral
assumptions
imply u’(y) > 0, u”(y) < 0, and u”‘(y) > 0.
H.R. Varian, Redistrihutioe
If no insurance
were available
the individual
taxation
53
choice of x would satisfy
u’(w -x) = Eu’(x + a).
If we assume u”‘(y)>O, Jensen’s inequality
implies Eu’(x+E)>u’(x),
and
hence x will be greater than w/2. Since no insurance is available, consumers
will be forced to self insure by oversaving.
If consumers
could purchase
insurance
against the cumulative
random events described by F, they would
clearly be able to eliminate
all individual
risk. If all individual
risk is
eliminated
the obvious individual,
and hence social, optimum
is to set x
= w/2.
On the other hand, there may be various
difficulties
with observing
realizations
of luck directly.
We assume that an external
observer
can
observe only measured
income r=x +E. Some part of measured income is
due to a decision by the individual
agent (x), and some part is due to
exogenous luck (E). For practical reasons, any compensation
must be based
on observed values of y=x SE rather than on E alone.
We let c(x + E) be the amount of consumption
allowed to a consumer with
observed income x + E. The objective of a social insurance program is to find
a consumption
function
that maximizes
expected
utility subject to two
constraints.
The first constraint
is the government
budget constraint:
the
receipts of the tax must, on the average, be equal to the amount payed out.
That is to say, the redistributive
program
must break even. The second
constraint
is that individual
consumers
choose x so as to maximize private
utility:
maxu(w-x)+Eu(c(x+a)).
X
(2)
This immediately
implies that the social optimum will be unattainable
by a
redistributive
tax based on observed income: if consumers were guaranteed
a
second period income of w/2, and a marginal tax rate of 1 on any income in
excess of w/2, they would choose to save nothing.
This is simply the
disincentive
effect of social insurance ~ a complete welfare state may result
in a significant incentive loss.
An optimal choice of x > 0 satisfies the lirst- and second-order
conditions:
u’(w-~)=Eu’(c(x+E))c’(x+E),
(3)
~“(W-~)+ELI”(C(X+E))C’(X+F)+EU’(C(X+E))C”(X+E)~O.
(4)
The first-order condition
is simply the statement
that the consumer
will
equate his expected marginal utility of income in each period. We note the
first two terms of the second-order
condition
are automatically
negative by
the assumption
of risk aversion.
The third
term is unfortunately
ambiguous
sign since it depends on the sign of c”(x +E). Our approach
will be to assume that the first-order
condition
completelv
describes
optimal choice of the consumer,
solve for the optimal c(r), and then
that such a c(y) satisfies the second-order
conditions.
As we shall see,
class of important
examples c”(y) will turn out to have the correct sign.’
Formulating
the above
discussion
algebraically,
we have the
insurance problem:
of
here
the
verify
for a
social
maxu(w-x)+Eu(c(x+s))
C(.LX
(5)
Ec(x+E)=x,
(6)
We investigate
this problem
in two stages.
First
we examine
determination
of the optimal linear tax, i.e. where c(y)=cq’+D.
Then
investigate the general problem of the optimal nonlinear
tax.’
the
we
3. The optimal linear tax
It is possible to derive the optimal linear tax by differentiating
(5) (7) and
manipulating
the resulting
conditions.
However,
a somewhat
clearer
derivation results if we proceed by a more indirect route.
The first step is to eliminate the governmental
budget constraint.
We write
after tax income as cq’+D. According to (6), and that fact that E&=0:
Thus,
D=(l-c)x.
(81
The next step is to use the individual
optimization
constraint
to define s
as a function of c, which we write as x(c).~ We now have a straightforward
‘The conditions
(3) and (4) are necessary and sufficient conditions for a local/): optimal choice
of x. It may happen that there are several solutions of (3) with only some of them being global
optima. Mirrlees (1975) has shown that this may cause diffkulties for the above technique.
The problem arises because the set of global optima - the set of solutions to problem (2) ~
can be a rather unpleasantly
structured
subset of the manifold described by (3). Mirrlees (1975)
suggests some techniques
to handle these diffkulties
in the finite dimensional
case, but there
seem to be no useful techniques available in the infinite dimensional
case. Hence, we adopt the
above approach
and simply assume that the first-order
condition
uniquely
determines
the
optimal choice.
2Note that when we restrict the optimal tax to be linear the second-order
conditions
(4) are
automatically
satisfied and the problem mentioned in footnote 1 does not arise.
3A standard
application
of the implicit function theorem shows that x(c) is a differentiable
function of c. This derivative is given in eq. (14).
H.R. Vurian, Redistributiw
unconstrained
maximization
problem
tuxcrtion
which determines
55
the optimal
maxu[w-x(c)]+Eu[c(x(c)+s)+(l-c)x(c)].
The derivative
z”(c)=
of the objective
function
tax:
(9)
is:
-u’(M.-x)x’(c)+Eu’(x+cE)
(10)
x[cx’(c)+X+&+(l--)x’(c)-x].
Applying the individual
first-order conditions
term and part of the second, leaving
z;‘(c)=Eu’(x+ce)(l
-c)x’(c)+Eu’(x+ce)~.
(3) we can eliminate
the first
(11)
We note first that Eu’(x +ce)s<O.
Since E&=0, this term is simply the
covariance
of marginal
utility and income; since these magnitudes
move in
opposite directions this covariance must be negative.
If we evaluate expression
(11) at c= 1 - the no-tax situation - the first
term drops out and we are left with a strictly negative value for u’(c). Hence
a decrease in c - an increase in the tax rate ~ must necessarily
increase
utility. This is simply the point that the first small imposition
of the tax
will result in no incentive loss, but will provide a small amount of insurance.
Therefore some amount of social insurance
is desirable, even in the face of
the incentive problem.
Setting U’(C) equal to zero we have a formula for the optimal tax:
In order to provide some intuition
for this expression
we imagine an
experiment of increasing c slightly to c + dc. On the one hand an increase in
c will result in more savings, since consumers
get to keep more of their
returns - savings go up by x’(c)dc. (Eq. (11) shows that x’(c) is positive at
the optimal c.) Because of this increase in savings, consumers will receive an
increase in the demogrant
D of (1 -c)x’(c)dc.
This in turn is multiplied
by
the marginal utility of second-period
income to determine the utility effect of
this change. On the other hand, the increase in c exposes the consumer
to
more risk. Instead
of facing a random
variable
CE, the consumer
faces
(c+~c)E. The expected marginal
utility due to this effect is Eu’(x +cs)~dc.
If the level of social insurance
is optimal, the sum of these two effects must
balance out to zero.
H.R. Varian, Redistributiue
56
Eq. (11) can be combined
C=
tuxution
with the individual
first-order
conditions
to give:
u’(w -x)x’(c)
uf(w-x)x’(c)-EU’(X+CE)E
(12)
This equation shows that the optimal marginal tax rate must be between 0
and 1.
We can derive an expression for the term x’(c) by differentiating
the tirstorder conditions
which define x. We find
u”(w-~)~‘(~)+~Eu”(x+cs)[x’(c)+~]+Eu’(x+cs)=O,
(13)
which can be rearranged
x’(c) =
to give
-EU”(X+CE)CE-EEu’(x+cE)
u”(w-x)+Eu”(x+c~)c
’
(14)
The denominator
of this expression is negative by virtue of risk aversion and
the numerator
will be negative if u”’ > 0. Hence, under an assumption
of
decreasing absolute risk aversion, x’(c)>O, a finding that is consistent
with
the equilibrium
condition described above.
4. An algebraic example
Let us
function
have the
l/2. The
apply .formula (11) to a simple problem involving a quadratic utility
and a discrete uniform distribution.
We let u(y) = by--$/2,
and let E
distribution
E = n with probability
l/2 and a= -n with probability
individual maximization
problem has first-order conditions :
-[b-(w-x)]+(c/2)[b-(x+cn)+b-(x-cn)]=O,
(15)
which can be solved for
x=w+bw)
.
lit
(16)
Thus,
x’(c)=
2b-w
(1
+c)2’
(17)
H.R. Varian, Redistrihutire
taxation
51
We also have
Eu’(x+cs)s=+[bUsing
(11) we obtain
(l_c)
(x+c~l)]n-+[b-
a formula
characterizing
(x-ccn)]n=cn’.
the optimal
(18)
c:
W-WI2
(1 +c)3
=m2,
Note that there will be a unique c that satisfies this equation. We can easily
examine the effects of parameter changes on the choice of the optimal c:
(i) as the size of the random income, n, rises, c must fall, so the optimal
tax rate rises;
(ii) as wealth increases, the left-hand side decreases, so c must decrease the optimal tax rate increases; and
(iii) as risk aversion increases, b must decrease, so the optimal tax must
increase.
Table
1
n
c
x
Utility
market
Utility
insurance
0.1
0.2
0.3
0.4
0.5
0.93
0.81
0.69
0.60
0.52
0.48
0.45
0.41
0.37
0.34
0.745
0.730
0.705
0.670
0.625
0.745
0.734
0.720
0.706
0.691
Table 1 presents some computations
of the optimal tax for various values
of the noise parameter n. In these computations
we set w = b = 1. The second
and third columns
of the table give the optimal values of c and x. The
optimal value of D is of course (1 -c)x;
this implies that the demogrant
as a
percentage of income is simply (1 -c). The last two columns give the utility
of the market (no intervention)
case, and the social insurance case. By way of
comparison,
the utility of the full insurance case is 0.750.
The main conclusion
to be drawn from this table is that in this example
the social insurance can capture only a small percentage of the gains possible
from full insurance, unless the randomness
in income is very extreme. In the
case where n =O.l, for example, second-period
income varies between 0.58
and 0.38, but second-period
consumption
is smoothed out to vary between
0.57 and 0.39 - not a very big change, and consequently
not a very big
contribution
to utility.
H.R.
58
Varian, Redistributiue
taxation
One could argue that this inability of the linear tax to capture much of the
potential
gains is due to the overly restrictive
assumption
of linearity.
A
more flexible tax structure may capture more of the gains. Accordingly,
in
the next section, we investigate
the problem
of determining
the optimal
nonlinear
tax.
5. The optimal nonlinear tax
In order to treat the nonlinear
tax problem, it is convenient
to rewrite (5)
(7) introducing
the density function
for E, which we denote by f(s). We
suppose that f(s) is strictly positive over the interior of the compact interval
[a, b] and that f(a)=f(b)=O.
Furthermore
we change the variable
of
integration
from E to ~=x+E,
so that the limits of integration
will now range
from y=x+a
to y=x+b.
For reasons that will be explained later, we have to be careful about the
domain of the utility function. We will suppose that utility is defined on the
non-negative
real line and that x and c(v) must lie in the compact region
05x Zw. With these changes eqs. (5)(7) become:
maxu(w-x)+Ju(c(y))f(y-x)dy
cC.1.x
subject
(19)
to
x-
&)f(y-x)dy=O,
(20)
(21)
It turns out to be convenient
to integrate
u’(w-x)+Ju(c(y))f’(y-x)dy=O.
eq. (21) by parts to get
(22)
Here we have used the fact that f(a)=f(b)=O
and multiplied
through by
- 1. The resulting
optimization
problem,
(19), (20), and (22) is a
straightforward
isoperimetric
calculus of variations
problem. It is, in fact, a
problem of the same form as that considered by Mirrlees (1974).
In that work, Mirrlees
used the Euler conditions
of the calculus
of
variations
to characterize
the optimal
solution
to a similar
problem.
However, in the same paper, Mirrlees also showed that in some cases no
optimal solution
would exist. It is worthwhile
reviewing his nonexistence
argument here, as it applies to the social insurance problem.
59
Let us choose consumption
schedules
of the form:
(23)
Here consumers
receive the first-best level of consumption
if their observed
income exceeds some target level a; if their observed income is less than I
they receive some (low) level of consumption
cl.
In order for this policy to be feasible it is sufficient to ensure that
consumers will, on the average, choose to save x=w/2. Mirrlees shows that if
utility is unbounded
below we can always choose c1 low enough so as to
induce this optimal level of savings.
Now consider what happens as we decrease the target level ~1, but adjust
c1 so as to keep x= w/2 and thus maintain
feasibility.
As cx declines
consumers have a higher and higher probability
of ending up with the firstbest level of consumption;
but if they are unlucky enough to end up with
observed income less than ~1,they will be more and more severely punished.
The net impact of these two effects on expected
utility is in general
ambiguous
but Mirrlees is able to show that in some cases total expected
utility will increase as the target level a decreases. Hence he can construct a
sequence of tax plans converging
to the first-best optimum,
but of course,
never actually reaching it.
In this way, we can approximate
as closely as we wish to the first-best
optimum, by imposing penalties (presumably
of great severity) on a small
proportion
of the population.
Although
these farmers suffer severely, there are so few of them that
their sufferings are outweighed
by the encouragement
their fate, or rather
the prospect of it, gives to farmers taking production
decisions. It seems
that models of this kind can in certain cases provide some justification
for
extreme punishment
of negligibly small groups [Mirrlees (1974)].
There are two important
problems raised by the Mirrlees result. The first is
the question
of the existence of an optimal tax. It seems clear that the
problem
here lies in M,irrlees’s assumption
of unbounded
utility. If one
imposes a bound on utility or consumption,
on grounds
of feasibility or
humanitarian
concern about ex post utility, the Mirrlees construction
can
only be carried so far - the unlucky consumers
can only be punished to a
certain degree. It seems possible that the imposition
of such a bound would
ensure the existence of an optimal tax.
However,
the Mirrlees’ example
raises another
question.
Even if the
optimal tax exists when consumption
is bounded,
how do we know that it
does not take the discontinuous
form considered
by Mirrlees? If it did take
such a form. the Euler conditions
used by Mirrlees would be irrelevant.
60
H.R. Vurian, Redistributive
taxation
In order to investigate
these questions
it is convenient
to pose the
optimization
problem
described
by (19X (20), and (22) as a problem
in
optimal control. To do this we create two dummy state variables defined as:
T(Y)=
R(y)=
j
[x-c(t)]f(t-xx)&,
1 [u’(Mi-x)+U(c(t))]f’(r-x)dt.
(25)
Xfll
Then we can write the relevant
maximization
problem
as:
maxu(w-x)+Su(c(y))f(y--x)dy
(26)
X.C(Y)
subject
to
T’(Y)
=Cx-c(Y)If(Y-X)t
(27)
R’(Y)
=[u’(w-x)+U(C(~))]f’(L’-X),
(28)
T(x+a) =T(x+b)=O,
(29)
R(x+a) =R(x+b)=O,
(30)
clc(y)jC;
oixjw,
(31)
This is a straightforward
problem
in optimal
control with the special
features that (1) the Hamiltonian
is independent
of the state variables, and
(2) the maximization
takes place over a parameter
x, as well as over the
function c(v). There are very general existence results for control problems
with feature (1) and in the appendix
I extend one of these results to the
parameterized
case. Hence, under quite weak assumptions
an optimal tax
schedule will exist.
A standard
application
of the maximum
principle
shows that at each
income
level y, c(y) will maximize
the Hamiltonian
function
over the
allowable values of c. We pose this maximization
problem here using 2 and ,U
to denote the adjoint variables:
~~~~x~(w--X)+~(C(Y))f(L.--X)+/.CX-c(r)lf(y-x)
+~Cu’(W--X)+~(C(~)lf’(L’-X)l
subject to
-C(Y)SC,
c(y)S;c.
(32)
H.R. Vurian, Redistributive
Letting p1 and p2 be the Kuhn-Tucker
constraints
we have the following conditions
choice of c(y) at each 4’:
61
tuxcction
multipliers
on the inequality
which characterize
the optimal
u’(c(y))[If(L’-x)+~~“(~-~)l=~~f():-x~-Pl
+pz,
PI 20,
c(L’)=c,
c<c(y)<C,
c(y)=C,
(34)
Pl ‘Pz
=o,
(35)
P*20.
We also have the first-order
Using the original formulation
(33)
(36)
condition
that x maximizes the Hamiltonian.
in (19))(21) this condition can be written as
j&Cl- ic’(Y)f(y-x)d+-A(x)=O.
(37)
Here we have used the fact that the derivative of the objective function must
vanish, owing to the individual
optimization
constraint,
and we have let d(x)
be the second derivative of individual
utility as given in eq. (4).
We first establish the following technical result.
Proposition.
If A#0
then c(y) is continuous
on the support
of f(a).
Proof
According to Fleming and Rishel (1975, theorem 6.1) we need only
establish that the Hamiltonian
has a unique maximizer at each value of y.
Referring to the KuhnTucker
condition
(33) we see that if [f(y-X)
+
$(J’-x)]
#O, the concavity
of u(c) guarantees
a unique solution to (33).
If the left-hand side of (33) is zero, then there are two cases:
(i) i_>O; then since f(y-x)>O
and ~~20,
we must have p1 >O and c=c;
and
(ii) 2 ~0; by analogous reasoning, we must have p2 > 0, and c = (_. Q.E.D.
This proposition
shows that as long as the aggregate income constraint
is
binding, the optimal tax will be continuous
on the support of E. Thus, for at
least some values of income, the optimal level of consumption
will be in the
interior
of [c, c] and the Euler conditions
derived by Mirrlees
will be
relevant. In the interior case, we have a very simple formula characterizing
c(y):
u’(c(y))
=
1
1 +/J&--2’)’
where h(~)=f’(~)/f’(~)
and !:=Ey=E(x+a)=x.
invert (38) to solve for C(J), up to the constants
Since u” ~0, we can directly
A and p.
62
H.R.
Vurian, Redistributive
taxution
We can use the information
contained
in eqs. (33)-(37) to provide some
qualitative
information
concerning
i and p. First, we note that at the modal
value of e ~ call it i=~+-r
we havef”(E*)=O, so
(39)
The complementary
slackness conditions
(34k(36) show that if the mode
occurs at a value of J: where c s c( 42)CC, then I will be positive. If c < c( J?) < (:
then /I is just the marginal utility of income at the modal value of E.
In order to provide an expression for 11, we rearrange (37) to get
(40)
The term in parentheses
is simply the average marginal
tax rate which we
would generally expect to be positive. Since i >O by (38) and d(x)<O, since
it is the second-order
condition
for individual
maximization,
we would
generally expect p to be negative.
In fact, it can be shown that p ~0 under the assumption
that h’(~)<0. To
see this differentiate (38) with respect to y to get:
- iph’(y - y)
“(4’)=LIC(C(J))(1
(40)
+/Lh(y-1’))2’
Suppose, prr ubsurdum, that ,LLL>O.Then (41) implies that c’(y)<O. This in
turn, by eq. (40) implies ,u <O. This contradiction
then establishes
three
important
facts: (1) ,U~0, (2) C’(Y)>O, and (3) the average marginal tax rate
is positive. (Analogous
facts were established
by Mirrlees (1974) for the
structure
of this problem
problem
he considered;
however, the additive
allows for a considerable
simplification
of the arguments4
)
The most interesting
behavior of the optimal tax formula given by (38) is
in the behavior
of the marginal
tax rate as a function of income. In the
optimal
taxation
literature
concerned
with the equity-efficiency
tradeoff
arising because of unobserved
differences of ability, it can be shown that the
marginal tax rate on the highest observed income must be zero. This effect
basically occurs because any tax schedule with a nonzero marginal tax rate
at the high end can be Pareto dominated
by one with a lower marginal tax
rate ~ the highest income person is made better off by lowering his marginal
tax rate and revenues available for redistribution
are unaffected. [See Brito
and Oakland
(1977), Cooter (1978). Phelps (1973), Sadka (1976), and Seade
(1977), among others.]
4Holmstriim
different
(1979)
conditions.
also
establishes
analogous
facts for a similar
problem
under
somewhat
H.R. Varian, Redistrihutiw
63
taxution
In the social insurance framework this effect is absent. Indeed, one would
suspect the marginal tax rate on the highest income person might be quite
large: if the only way to become a millionaire
is to be lucky, there should be
very small incentive losses from taxing a million dollar income at a high rate
- and there are substantial
insurance
gains since one can then subsidize
million dollar occurrences of bad luck.
Let us examine this argument
more rigorously.
First, if the C constraint
is
binding at high levels of income, the marginal tax rate is clearly one, so we
concentrate
on the interior case. Differentiating
(38) we have:
~“(c(Y)k’(.Y)=
(1
Using
- 1”
+ph(y))2
Ph’(4’1.
(42)
(38) and rearranging:
(43)
Letting p(c(y))= -u”(c’(~))/u’(c(~))
risk aversion, we have
be the Arrow-Pratt
measure
of absolute
This equation shows that, ‘other things being equal’, a higher degree of risk
aversion implies a higher marginal
tax rate. (Of course other things can
never be equal since they also depend on the degree of risk aversion.)
Differentiating
(44) once more we have
The term in front of the brackets will generally be
to assume h’z0, h”zO the first term inside the
which suggests a declining marginal
consumption
marginal
tax rate. The second
term, however,
direction.
If absolute risk aversion decreases with
will be positive and the overall sign of c” will be
will imply c” <O, but this seems an implausible
grounds. However, it does seem clear that if risk
too rapidly, then c”(y) will be negative and our
supported.
negative. If we are willing
brackets will be positive
rate - i.e. an increasing
works in the opposite
income, the second term
ambiguous.
Clearly p’ $0
assumption
on a priori
aversion does not decline
earlier intuition
will be
A nice example to illustrate
these considerations
relative risk aversion utility function and a normal
~(c)=c~-~/(l
-p), (33) implies that
is that of a constant
distribution
on E.’ Since
(46)
or
c(y)=[B+Ah(y)]l’P,
(47)
where B > 0, A < 0, and h’(y) < 0. In this case p’ = 0 and the first term in (44)
always dominates so that c”(y) < 0.
If E is normally distributed
with
f(~) = k epe212,
(48)
h(E)=+
(49)
we have
--E,
&
so that the optimal
tax takes the form:
c,
c(y) =
if
1
cz[B-Aq.]“P
[B - AL’]“P
(;
if
(50)
F~[B-A~]“P.
If p= 1, a linear tax is optimal. As risk aversion increases, a more progressive
tax is desirable, exactly as our intuition suggested.
The case with p= 1 is an especially
simple case in which to calculate
optimal taxes. Table 2 presents an example of one such computation
with
w =40, CT=2 and 3. As before, the columns
labelled ‘market’, ‘insurance’,
and ‘full’ give the utility from each arrangement.
Note the relatively low marginal tax rate. In the case where g= 3, income
would lie in the two standard
deviation
range 1426, but the marginal tax
rate would be only 4 percent. The demogrant
would also be 4 percent of
first-period income. The partial insurance offers a ‘small’ improvement
on the
market solution,
but does not capture much of the potential
gains from
complete insurance.
50f course, for an optimal
compact support.
tax to exist we should
truncate
the distribution
on c so that it has
65
Table 2
0
c
x
Market
Insurance
Full
2
3
0.98
0.96
19.897
19.815
5.9863
5.9799
5.9864
5.9804
5.9914
5.9914
6. A curious example
In some circumstances
the optimal tax may take a rather peculiar form.
Suppose,
for example,
that there are occurrences
of E which reveal
unambiguously
the ex ante actions of the consumers.
For example, suppose c
has support (u, h) and w/2 fu >O. Then if one observes measured income J
in the interval w/2 + CI> y > 0, one can be certain that the consumer saved less
than w/2 in the first period. In this circumstance
it makes sense to choose a
consumption
schedule that would give the consumer
a very low level of
consumption
in this region; so low, in fact, that the consumer would always
save x=w/2
so as to avoid falling into this region. In this way, we can
support the first-best optimal solution.6
The mathematical
details go as follows. As in section 5, let us choose a
policy of the form:
c(x+Fi)=q,
w/2,
if
if
x+E<w/2+a,
x+F~w/~+~.
We want to choose c1 small enough so that the utility from choosing x = w/2
exceeds the utility from choosing any other choice of x. That is, we choose c1
so that
2u(w/2)>maxtl(w-x)+u(w/2)
X
+[u(cl)-4w/2)]F(w/2+a-x),
(51)
where F( .) is the cumulative
distribution
function_ of E.
For fixed cl, the right-hand
side of this inequality
is bounded above, so a
maximum exists. As long as u(cl) can be made sufficiently small, we can be
assured of satisfying inequality
(51) and thereby being able to support the
first-best optimal pattern of consumption.
‘Harris and Raviv (1978a.b) have independently
noted
agent problems often exists under similar circumstances.
that a first-best
solution
in principal
7. Summary
and conclusions
If income contains
a random component
then a system of redistributive
taxation
will contribute
to reducing the variance of after-tax income. The
design of an optimal
program
of redistributive
taxation
must weigh the
benefits from this ‘social insurance’ against the deadweight costs of a tax on
transactions.
In this paper we have examined the determination
of such an
optima1 tax in a simple choice theoretic framework.
In the case of linear taxation
the optimal
tax can be derived from a
relatively
simple and intuitive
formula. The nonlinear
tax involves a less
intuitive but surprisingly
explicit formula. We have shown that if one bounds
the allowable values of consumption
the optimal tax will always exist and
generally be continuous
as a function of observed income. Under reasonable
assumptions
about the distribution
of the randomness
in income, the optima1
tax will be increasing
in income and the average marginal tax rate will be
positive.
The optimal
tax may exhibit an increasing
marginal
tax rate,
depending on the nature of the stochastic influences on income.
Finally
it should be remarked
that the analysis of this paper can be
applied
to a number
of other problems
involving
moral
hazard
and
insurance.
In particular,
the existence and continuity
results can be easily
generalized to the principallagent
problems considered by Harris and Raviv
(1978a, b), Holmstrom
(1979) and Shave11 (1975).
Appendix
Existence of an optimal solution
Here we state an existence theorem for a class of optimal
that is referred to in the text. Consider the control problem:
subject
control
problems
maxj! h,(t,u(t),z)dt
U(.),-_t 0
(A.1)
dx
z=h(t,
(A.21
to
u(t), z),
x(tl)=xl(z),
x(b)=.%(z);
u(t)
z
x(t)
in
in
8,
Z,
in
a compact
a compact
subset of R”,
subset of R”,
R”,
x0, x,, h, and h continuously
differentiable.
Here Y is the vector of state variables, u is the vector of control variables,
and z is a vector of parameters.
We assume that a feasible solution exists.
For fixed Z, this is a standard control problem with the special feature that
11, and /I are independent
of the state variables. Theorem 6.3 of Berkovitz
to this control
(1974, p. 165) shows that for fixed z an optimal solution
problem exists. We wish to extend this argument
to the parameterized
case.
First we introduce some notation.
Given an admissable
control u( .) and a parameter
Z. consider the system
of differential equations given by
dr
d’r =Iro(t, u(t), Z),
(A.3)
The ‘attainable
set’ (for fixed 2) is defined to be the set of states
system at t, as u( .) ranges over the set of admissible controls.
&=((?:x)ER
n+l :?‘=!.(t,);
where J,( ) and s(
of this
x=x(r,),
) solve (A.3) for some admissible
control
[I( ), )
Note that y=~(tr ) is simply the value of the objective function in problem
(A.1). Berkovitz (1974, theorem 6.3) shows that K, is compact. so clearly A’
achieves a maximum value on K,. Hence an optimal solution to the control
problem with fixed z and no boundary
constraint
at t, exists.
For the parameterized
problem the analogous attainable
set is:
This is clearly compact since it is the product of two compact sets.
Finally, we incorporate
the additional
boundary
constraint
by defining:
This set is compact
since it is a closed subset of K; it is nonempty
hypothesis.
Hence 4’ attains a maximum
on K’ and a solution to (A.l)
(A.2) therefore exists.
by
and
68
H.R. Varian, Redistributiw
taxutim
References
Atkinson,
A. and J. Stiglitr, 1976, The design of tax structure:
Direct versus indirect taxation,
Journal of Public Economics 6, 55576.
Berkovitz, L., 1974, Optimal control theory (Springer-Verlag,
New York).
Beveridge, W., 1942, Social insurance and allied services (Macmillan, New York).
Brito, D. and W. Oakland,
1977, Some properties
of the optimal income tax, International
Economic Review 407 -424.
Coe, Richard, 1977, Dependence
and poverty in the short and long run, Survey Research Center.
University of Michigan.
Diamond, P., 1975, A many person Ramsey tax rule, Journal of Public Economics 4, 3355342.
Diamond,
P. and J. Mirrlees, 1971, Optimal
taxation
and public production
IIII, American
Economic Review 61, &27, 261-278.
Diamond,
P. and J. Mirrlees, 1978, A model of social insurance with variable retirement,
M.I.T.
Working Paper no. 210.
Diamond,
P.. J. Helms and J. Mirrlees, 1978, Optimal
taxation
in a stochastic
economy,
A
Cobb-Douglas
example, M.I.T. Working Paper no. 217.
Fleming, W. and R. Rishel. 1975, Deterministic
and stochastic optimal control (Springer-Verlag,
New York).
Harris, M. and A. Raviv, 1978a. Some results on incentive contracts,
American
Economic
Review 68, 20 -30.
Harris, M. and A. Raviv, 1978b, Optimal incentive contracts with imperfect Information,
Journal
of Economic Theory 20, 231-259.
Holmstrom,
B., 1979, Moral hazard and observability,
Bell Journal of Ecc .)mics 10, 74-91.
Jencks, C., 1972, Inequality (Basic Books, New York).
Lillard, L. and R. Willis, 1978. Dynamic aspects of earnings mobility. Econometrica
46, 985
1012.
Mirrlees, J., 1971, An exploration
in the theory of optimum
income taxation,
Review of
Economic Studies 38, 175 208.
information,
and uncertainty,
in: Balch,
Mirrlees,
J., 1974, Notes on welfare economics,
McFadden
and Wu, Essays on economic
behavior
under uncertainty
(North-Holland,
Amsterdam).
Mirrlees, J., 1975, On moral hazard and the theory of unobservable
behavior, Oxford University,
mimeo.
Phelps, E., 1973, Taxation of wage income for economic justice, Quarterly Journal of Economics
87, 331-354.
Ross, S., 1979, quoted in Family Weekly, June 24.
Sadka, E., 1976, On income distribution,
incentive effects, and optimal income taxation, Review
of Economic Studies 43, 261-267.
Sandmo,
A., 1976, Optimal
taxation
-- An introduction
to the literature,
Journal
of public
economics 6, 37754.
Geade, J., 1977, On the shape of optima1 tax schedules, Journal of Public Economics 7, 203 -236.
,‘lavell, S., 1975, Risk sharing and incentives
in the principal
and agent relationship,
Bell
tournal of Economics 10, 55 73.